A Preliminary Numerical Study to Compare the Physical Method and Machine Learning Methods Applied to GPR Data for Underground Utility Network Characterization

Abstract

In the field of geophysics and civil engineering applications, ground penetrating radar (GPR) technology has become one of the emerging non-destructive testing (NDT) methods thanks to its ability to perform tests without damaging structures. However, NDT applications, such as concrete rebar assessments, utility network surveys or the precise localization of embedded cylindrical pipes still remain challenging. The inversion of geometric parameters, such as depth and radius of embedded cylindrical pipes, as well as the dielectric parameters of its surrounding material, is of great importance for preventive measures and quality control. Furthermore, the precise localization is mandatory for critical underground utility networks, such as gas, power and water lines. In this context, innovative signal processing techniques associated with GPR are capable of performing physical and geometric characterization tasks. This paper evaluates the performance of a supervised machine learning and ray-based methods on GPR data. Support vector machines (SVM) classification, support vector machine regression (SVR) and ray-based methods are all used to correlate information about the radius and depth of embedded pipes with the velocity of stratified media in various numerical configurations. The approach is based on the hyperbola trace emerging in a set of B-scans, given that the shape of the hyperbola varies greatly with pipe depth and radius as well as with velocity of the medium. According to the ray-based method, an inversion of the wave velocity and pipe radius is performed by applying an appropriate nonlinear least mean squares inversion technique. Feature selection within machine learning models is also implemented on the information chosen from observed hyperbola travel times. Simulated data are obtained by means of the finite-difference time-domain (FDTD) method with the 2D numerical tool GprMax. The study is carried out on mono-static, ground-coupled GPR datasets. The preliminary study showed that the proposed machine learning methods outperforms the ray-based method for estimating radius, depth and velocity. SVR, for instance, calculates depth and radius values with mean absolute relative errors of 0.39% and 6.3%, respectively, with regard to the ground truth. A parametric comparison of the aforementioned methodologies is also included in the performance analysis in terms of relative error.